Robodebt Billed 433,000 People. The Formula Wasn't the Real Flaw.

Plain English Summary

Australia sent debt notices to about 433,000 people between 2015 and 2019. The notices rested on a shortcut: the agency took a person's yearly income, split it evenly across the year, and guessed what they were paid each fortnight. The shortcut was illegal. But I argue a second flaw did more harm. The person who got the notice had to prove it was wrong. The agency did not have to prove it was right. A rule that puts the proof on the agency would have stopped most of the damage, even with the same formula.

The question

Between July 2015 and November 2019, the Australian government raised about $1.73 billion in debts against about 433,000 people, according to the Royal Commission's report as summarised by the Australian Policy Online library [1]. About 381,000 of them were pursued through private debt collectors, and almost $752 million was repaid [1].

The standard story is that the formula was bad. The scheme used "income averaging" (dividing annual income from tax records evenly across fortnights) to decide what a person earned in each fortnight. The Royal Commission found the scheme was "neither fair nor legal and it made many people feel like criminals" [2].

My question is narrower. Suppose the formula had stayed exactly as it was. Which single rule would have done the most to prevent the harm? I think the answer is a rule about who must prove the debt, not a rule about the arithmetic.

Data and where it came from

I could not open the full report. The Royal Commission's own site timed out, and the PDF mirror returned an access error. So every figure and quote below comes from summaries of the report, a Federal Court case summary, an academic law blog and practitioner commentary. If you need a page citation, check the report itself.

What I did read:

• The Federal Court case summary of the class action settlement, approved on 11 June 2021 [3].
• Law commentary from 2018 on what the Social Security Act requires [4].
• Practitioner coverage of the Royal Commission's findings [5][6].
• A search-result summary of the Commonwealth Ombudsman's April 2017 report [7].
• A search-result summary of Recommendation 17.1 [8].

Method

I use one test. For each failure, I ask whether it would have survived if the burden of proof (the party who must supply evidence to win) sat with the agency. If it would not survive, the burden rule is the stronger safeguard. If it would survive, the burden rule is not enough.

I did this without the Lab. The arithmetic below is simple division of published figures, and you can repeat it with a calculator.

The burden decides who loses when the facts are unclear. It matters most here because the facts were often unclear: people were being asked about pay from years before.

What the law required

The Federal Court recorded that "the Commonwealth conceded, and the court found, that it did not have a proper legal basis to raise, demand or recover asserted debts based on income averaging from ATO data" [3]. Justice Murphy called the scheme "a shameful chapter in the administration of the Commonwealth social security system and a massive failure of public administration" [3].

Note the framing: "did not have a proper legal basis." That is a statement about proof, not only about arithmetic.

A 2018 law analysis explains why. Section 1222A(a) of the Social Security Act 1991 says a debt exists only if another provision creates it, and section 1223 requires the agency to establish a difference between what was paid and what the person was entitled to [4]. The same author argued that the agency "bears the onus of establishing the existence and size of any debt" [4]. One lawyer quoted in the Law Society Journal said you cannot reach a state of satisfaction about a past fortnight "by looking at an annual figure for income and averaging it out" [5].

So there were two defects, and they are different.

1. A wrong estimator. Averaging produces a number that is not the person's pay in that fortnight.
2. A reversed burden. The scheme treated the average as a finding and told the person to disprove it.

A community legal centre summary puts the second defect in one line: "This effectively reversed the onus of proof from a powerful government agency to individual social security recipients" [6]. Another lawyer quoted in the Law Society Journal said it "effectively shifted the onus for proving whether debt existed from Centrelink onto social security recipients" [5].

Result: the formula was a fine tip-off and a bad verdict

Here is my central claim. Income averaging, used as a screen, is not unlawful. A mismatch between yearly tax data and fortnightly reports is a reasonable reason to ask a question. The same formula used as a verdict is unlawful.

The harm came from the step between those two uses. The agency sent a notice, and if the person did not answer, it raised the debt. The Royal Commission's findings, as reported, say no regard was given to how unreasonable it was to ask people to establish their earnings for periods going back as long as five years [5][6]. I could not confirm that wording against the report, so I do not put it in quotation marks.

Three numbers show the scale, using the published figures [1]:

Quantity	Calculation	Result
Average debt raised	$1.73 billion / 433,000	about $3,995
Share of debtors pursued by collectors	381,000 / 433,000	about 88%
Average repaid per pursued person	$752 million / 381,000	about $1,974

These are averages of rounded published figures. They hide the spread, and I do not have the distribution. But they say something. Roughly nine in ten people with a debt were sent to collectors. For most people, the system worked as a debt collection pipeline, not as a question.

The Ombudsman's 2017 report matters here, but I must be careful. A figure of about 20 percent was widely called an error rate. The summary I saw says it counts customers who got a letter, explained themselves, and then had no debt [7]. That is not a false positive rate of the formula. It is the share of people who responded and could supply an answer. It shows the burden working as a filter on people who could respond. It says nothing about people who did not.

That distinction is the heart of my argument. The people who could not respond were not people with correct debts. They were people the system could not hear.

Who could appeal?

Name the decision maker. In the scheme, it was an automated process inside the agency, run on tax data. The person with the debt could ask for an explanation or review, but only after the notice and only with their own records.

Who can appeal? Anyone who had the records, the time, and the knowledge that they could. A summary of the Royal Commission's findings notes numerous first-tier tribunal decisions striking down income averaging, yet the scheme continued [6]. The appeal path worked for a few people and never fed back into the rule.

I made a related point in the Michigan MiDAS post, where the state seized a refund before the appeal ended. I extend it here: in Michigan the state collected first. In Australia the state asserted first, and the person had to answer. In the Dutch benefits post I argued that human reviewers who cannot see why a flag was raised add little. The Australian case adds a harder point: reviewers who can see the flag still add little if the rule says the person must disprove it.

Sensitivity: what would overturn my claim

The assumption that moves my result most is this: what share of the 433,000 averaged debts matched a real overpayment?

The settlement wiped out all debts raised in whole or part through averaging: about $1.763 billion in withdrawn debts and about $751 million refunded [3]. So the debt total is an upper bound on money not properly owed, not a measure of money falsely billed. Some people may really have been overpaid. The Commonwealth lacked a proper legal basis to say so.

If many debts were real, a burden rule would still have let the agency recover them, but only after it produced employer pay records. If few were real, a burden rule would have stopped nearly everything. I do not have the share. The sources I read do not give it.

Here is scenario arithmetic, not data. Let $p$ be the share of the 433,000 debts that match real overpayments. Assume the agency can prove all of those and none of the others. Then a burden rule would produce about $433{,}000 \times p$ lawful debts.

• If $p = 0.2$, about 86,600 debts survive.
• If $p = 0.5$, about 216,500 survive.

Both are a large cut from 433,000. The conclusion that the burden rule is the stronger safeguard holds across that range. It fails only if $p$ is close to 1, which the Commonwealth's concession and the tribunal decisions make hard to believe [3][6].

The second sensitivity is the cost of proof. Under my rule the agency must get employer records for each flagged person. That costs money and time, and the agency would recover less, later. I do not price this. I have a habit of leaving that gap, so read the rule as a proposal, not a finding.

The third point is that the Commission asked for more than a burden rule. Recommendation 17.1, as summarised, calls for a clear path to review for people affected by automated decisions, plain-language notice that automation is used, and publication of business rules and algorithms so experts can scrutinise them [8]. I think those are good. But a path to review that starts after the debt exists is the path that failed. A burden rule acts before the debt.

The rule

One rule: no automated flag may create or increase a debt. A debt needs a record for the specific period, produced by the agency, before any demand for payment.

Default if the institution does not exist: if a body has no way to get that record, it may send a question but not a demand. The flag stays a lead.

How it could fail: the agency can satisfy "a record" with a cheap stamp, such as an automated request to an employer that returns nothing and is counted as proof of effort. Or the burden can creep back through a clause that says the person must take "reasonable steps" to supply what the agency lacks. A tired official applying that rule at scale will drift toward the second reading. So the rule needs a second line: if the agency cannot produce the record, the debt is zero, and that outcome is logged and reported.

What would change my mind: a published breakdown showing that most of the 433,000 debts matched real overpayments, or evidence that the agency could not have obtained the records at any reasonable cost. If either holds, the burden rule would still be right on principle but too expensive to be the main safeguard, and I would say so.